ByteDance lifts 2026 AI capex to 200 billion yuan

ByteDance lifts 2026 AI capex by 25 percent

ByteDance, the parent of TikTok and Douyin, plans to raise its 2026 AI infrastructure spending by roughly 25 percent to 200 billion yuan (about US$29.4 billion). Bloomberg and the South China Morning Post reported the change on 9 May 2026, citing people familiar with the matter. Late in 2025 the company had communicated a 2026 capex target of 160 billion yuan.

What the increase actually covers

According to the reports, the additional money goes into three buckets: GPUs and accelerators for training and inference; memory chips, whose prices have risen sharply; and additional data-center capacity. Bloomberg writes that a growing share of the budget is being steered toward domestically produced AI chips such as Huawei Ascend, in order to reduce exposure to US export controls.

Memory prices as a driver

Sustained high prices for HBM3 and HBM4 memory are cited as a key reason for the lift. Memory has become, alongside GPUs, the second major bottleneck in building large training clusters.

Shift toward Chinese accelerators

ByteDance has leaned heavily on NVIDIA in recent years. The growing share of Chinese chips is a direct response to the tightened US export rules in force since 2024 and to Beijing's push for technological self-sufficiency.

Comparison with US hyperscalers

The US$29.4 billion figure is still well below the major US hyperscalers. Amazon has guided to roughly $200 billion of 2026 capex, Alphabet and Microsoft to about $190 billion each, and Meta to $145 billion (per the companies' own Q1 2026 disclosures). ByteDance is therefore China's largest private AI spender, but trails the US cluster on an individual basis.

Why this matters

The 25 percent jump shows that the AI capex wave is not subsiding, neither in the US nor in China. For memory and compute-chip suppliers it means stable to growing demand and continued tight allocations. For European observers, the relevant point is that China is deliberately reducing reliance on Western accelerators, which further accelerates the US–China tech decoupling. Mid-market firms in Germany, Austria and Switzerland feel that indirectly: through GPU prices, longer lead times for high-end hardware, and through the question of which cloud region can run a model in a safe and compliant way.

In plain language

Imagine a big restaurant had already ordered 160 sacks of flour for 2026 because it plans to bake a lot more bread. Now the chef says: make it 200 sacks. And some of them we no longer buy from the old supplier overseas — we buy from the local mill in the next village, because the borders have got shakier. ByteDance is buying the same kind of thing for AI: more raw ingredients, and more of them sourced from its own region.

A practical example

A 600-employee machine builder in Baden-Württemberg plans to build its own predictive-maintenance AI platform in 2026. Instead of buying eight NVIDIA H200s outright, it explores a mixed model: four H200s on premises for sensitive training runs and additional inference in a European cloud. As long as ByteDance, Tencent and Alibaba keep stockpiling hardware in parallel, mid-market buyers should plan with lead times of 8 to 12 months and price premiums of 10 to 20 percent versus 2024. Contracts with cloud providers are worth pinning down on explicit hardware availability, generation and region.

Scope and limits

The 200 billion yuan figure comes from anonymous sources cited by Bloomberg and SCMP. ByteDance has neither officially confirmed nor denied it. The exact split between NVIDIA, Huawei Ascend and other accelerators is not public. Capex plans are targets, not actuals: real deliveries depend on sanctions, currency markets and Chinese domestic chip production. Strategic decisions should therefore not rest on these capex estimates alone.

SEO and GEO keywords

ByteDance, TikTok, AI capex 2026, AI infrastructure, China AI, Huawei Ascend, NVIDIA H200, HBM3, AI chips, hyperscaler, US-China tech, AI investments